Add custom implementation of half float

CMakeLists.txt

@@ -54,6 +54,7 @@ set(TENMON_SRC
     statusbar.cpp statusbar.h
     stfslider.cpp stfslider.h
     stretchtoolbar.cpp stretchtoolbar.h
+    tfloat16.h
 )
 
 qt_add_resources(TENMON_SRC resources/resources.qrc)

rawimage.cpp

@@ -1,6 +1,7 @@
 #include "rawimage.h"
 #include <cstring>
 #include <lcms2.h>
+#include <algorithm>
 #ifndef NO_QT
 #include <QDebug>
 #include <QElapsedTimer>
@@ -8,9 +9,16 @@
 #include <QColorSpace>
 using F16 = qfloat16;
 #else
-#include <algorithm>
+#define __STDC_WANT_IEC_60559_TYPES_EXT__
+#include <float.h>
+#ifdef FLT16_MAX
 using F16 = _Float16;
-#endif
+#else
+#include "tfloat16.h"
+using F16 = TFloat16;// this is only for MXE
+#endif // FLT16_MAX
+
+#endif // NO_QT
 
 int THUMB_SIZE = 128;
 int THUMB_SIZE_BORDER = 138;
@@ -464,13 +472,25 @@ void RawImage::convertToThumbnail()
 
                 if(m_channels == 1)
                 {
-                    out[idx] = out[idx + 1] = out[idx + 2] = (F16)(in[idx2] * scale);
+                    if(scale == 1.0f)
+                        out[idx] = out[idx + 1] = out[idx + 2] = (F16)(in[idx2]);
+                    else
+                        out[idx] = out[idx + 1] = out[idx + 2] = (F16)(in[idx2] * scale);
                 }
                 else
                 {
-                    out[idx]     = (F16)(in[idx2] * scale);
+                    if(scale == 1.0f)
-                    out[idx + 1] = (F16)(in[idx2 + 1] * scale);
+                    {
-                    out[idx + 2] = (F16)(in[idx2 + 2] * scale);
+                        out[idx]     = (F16)(in[idx2]);
+                        out[idx + 1] = (F16)(in[idx2 + 1]);
+                        out[idx + 2] = (F16)(in[idx2 + 2]);
+                    }
+                    else
+                    {
+                        out[idx]     = (F16)(in[idx2] * scale);
+                        out[idx + 1] = (F16)(in[idx2 + 1] * scale);
+                        out[idx + 2] = (F16)(in[idx2 + 2] * scale);
+                    }
                 }
                 out[idx + 3] = (F16)1.0f;
             }
@@ -546,9 +566,9 @@ void convertType2(size_t size, const T *src, U *dst)
 
     if constexpr(std::is_integral_v<T> && (std::is_floating_point_v<U> || std::is_same_v<U, F16>))
     {
-        U scale = (U)(1.0 / (double)std::numeric_limits<T>::max());
+        float scale = (float)(1.0 / (double)std::numeric_limits<T>::max());
         for(size_t i = 0; i < size; i++)
-            dst[i] = (U)src[i] * scale;
+            dst[i] = (U)(src[i] * scale);
     }
 }
 

tfloat16.h

@@ -0,0 +1,61 @@
+#ifndef TFLOAT16_H
+#define TFLOAT16_H
+
+// crude implementation of float16 for platforms that do not support _Float16
+
+#include <stdint.h>
+
+class TFloat16
+{
+    uint16_t b16;
+public:
+    TFloat16(){ b16 = 0; }
+    explicit inline TFloat16(float f)
+    {
+        uint32_t i = *reinterpret_cast<uint32_t*>(&f);
+        uint32_t sign = (i >> 16) & 0x8000;
+        uint32_t exp = (i >> 23) & 0xff;
+        uint32_t mantisa = (i & 0x7fffff) >> 13;
+
+        b16 = 0;
+        if(exp < 111)
+        {
+            // do nothing it map to 0
+        }
+        else if(exp == 111)
+        {
+            b16 |= sign;
+            b16 |= mantisa;
+        }
+        else if(exp > 142)
+        {
+            b16 = 0x7c00;// inf
+            b16 |= sign;
+        }
+        else
+        {
+            b16 |= sign;
+            b16 |= (exp - 112) << 10;
+            b16 |= mantisa;
+        }
+    }
+    friend TFloat16 operator*(TFloat16 a, TFloat16 b)
+    {
+        return TFloat16(static_cast<float>(a) * static_cast<float>(b));
+    }
+    operator float() const
+    {
+        uint32_t i = 0;
+        uint32_t sign = b16 & 0x8000;
+        uint32_t exp = (b16 & 0x7c00) >> 10;
+        if(b16)
+        {
+            i |= sign << 16;
+            i |= (exp + 112) << 23;
+            i |= (b16 & 0x3ff) << 13;
+        }
+        return *reinterpret_cast<float*>(&i);
+    }
+};
+
+#endif // TFLOAT16_H